應用電腦立體視覺建立虛擬實境地理資訊系統

Abstract

本研究利用電腦立體視覺建立環場的虛擬場景，並配合前人的研究，建立一個虛擬實境地理資訊系統。在傳統的電腦立體視覺中，大多僅針對某一特定目標進行重建，本研究將此方法加以延伸，藉由前人架設好的儀器，以旋轉CCD攝影機的方式拍攝目標地點四周的多組立體視覺影像，依此計算出數個三維模型。我們主要採用的比對方法是改良的Graph Cut方法，再藉由Rank演算法做輔助，並利用像差影像做檢視和操作的媒介，提供即時回饋的後處理方式，能夠簡單並快速地修正比對錯誤的部分。得到所需的三維模型後，我們使用自行發展的演算法將模型轉換成圓柱座標，然後依據其真實地理位置計算相重疊的部分，依據重疊位置之角度給予兩個模型不同的權重並疊加起來，組成一個環場的三維虛擬場景。由於計算得到的資料量相當龐大，因此必須進行資料減量，刪除不必要的稜邊，減少多邊形數目，以增進顯示效率，節省儲存空間。這個部份則是利用Memoryless Simplification演算法，藉由GTS函式庫實做完成，在減量至80%的情況下，與原始模型的誤差為13.01%。經過上述程序處理後的資料，將配合前人的研究，存入地理資訊系統資料庫中，使用者能夠藉由我們自行開發的程式查詢電子地圖上各點的地理資訊，並且同時在虛擬場景中遊覽。我們藉由事前手動設定資料與執行時自動偵測，可以順暢地切換場景，並阻止使用者移動到場景以外的地方，讓使用者在數個場景中遊走時能夠平順自然。

In this paper, we built a virtual panorama scene based on computational stereo, and combined it with GIS which is developed by former studies to be a virtual reality GIS. Traditional computational stereo problems confer on only one stereo pair, this study expends it to a panorama case. We rotate two CCD cameras to grab several images around a specific place and then generate 3D models by these stereo pairs. Our main matching algorithm is the improved Graph Cut method, and the Rank method is used as a fast solution. We also provide a post-processing method that is performed depending on the disparity image and has real-time feedback so that we can easily correct bad points deriving from false detections. After getting all models what we need, we merge these models into a 3D virtual panorama scene. We use our own algorithm to translate models into cylinder coordinate system, find the overlapped region, assign weights according to the spatial locations, and then blend them. Since the models generated by computational stereo always have huge size, we use the memoryless simplification to simplify them for fast rendering and saving memories. The error between original model and 80% simplified model is 13.01%. This part is practiced by the GTS library. After doing all of the above, we save the virtual scenes into the GIS developed by former studies, so the users can query geometrical information and visit the virtual scene at the same time in our program. According to the manual preprocessing, we can auto detect at runtime to change scenes or prevent user’s moving to hold he in the scene, so the user can explore the scenes smoothly.